Model-based damage identification with simulated transmittance deviations and deep learning classification

Abstract

Damage detection and identification is one of the main tasks in vibration based Structural Health Monitoring (SHM). The robustness of such SHM applications depends among others on the amount and quality of data that can be acquired. Model-based SHM methods may offer such data in unlimited numbers by simulating different structural states; however the main drawback remains the accuracy of models especially for small damages and early detection scenarios. In the present work, a method is presented where SHM data is generated though Finite Element (FE) models, simulating transmittance deviations from reference healthy states. The method is tested on a Carbon Fiber Reinforced Polymer truss for multiple damage scenarios of relatively realistic and small magnitude, affecting different truss members. The transmittance deviations for each scenario are approximated in the FE model by reducing the stiffness of the corresponding components simulating in parallel different uncertainties, resulting in a rich training dataset. The simulated data is finally passed to a Deep Learning (DL) classifier which is later validated on the experimental damages. The dataset is proven to provide to the DL classifier the appropriate information to generalize on the experimental states and the method has potential to contribute to model-based SHM applications. The numerical to experimental generalization is proven to depend on the uncertainty simulation of various model parameters.